dsh-vision-plugin

The deepseek-v4 family in DeepSeek Harness is strictly text-only: paste a screenshot, a table photo, or a chart into the chat and the model simply cannot see it. dsh-vision-plugin bridges that gap. It intercepts any image that would land in a text-only session, routes it through a vision model to produce a text transcription, and feeds that description back so the main model answers as naturally as if it had native vision.

Three layers of capability come with it. The lightest is a zero-ceremony path: drop an image into the chat and any model that has declared image capability can answer immediately, no special syntax required. The middle layer is the vision_analyze tool, which the model can invoke on its own to inspect a local image file, with an optional question and an optional vision-model override. The outermost layer is a Vision Model page in the DSH settings panel where you pick the default vision model, inspect the active route, and have your choice persisted and restored across restarts. Under the hood, transcriptions are cached by image-and-question pair so follow-up turns are free, hung vision calls time out after two minutes, and a total failure triggers one automatic retry with a fallback model before the conversation notes the image is unavailable and moves on.

This plugin is for anyone running DeepSeek Harness who works primarily with text-only models but regularly needs to make sense of screenshots, table photos, or charts. As long as a vision model is available in your provider configuration, installing the plugin lets every declared text-only model read images seamlessly without switching your main model.